Solomon v1.1
LoRA adapter and trained answer heads for Qwen3.8-27B. A document is read once into a reusable state; yes/no, single-choice, ordered-choice and multi-label questions then get probabilities read from letter logits, with optional ranked sentence pointers. No text is generated.
A general document-decision model; the card makes no clinical claims. Doccy's own panel is 802 questions over 54 public documents, with labels made by AI and never reviewed by a human. The v1.1 gain over v1.0 (+3.4 points) is not statistically established: the 95% document-bootstrap interval includes zero. Probabilities are served unscaled and miss the maker's own 0.03 ECE target for yes/no and ordered questions; answers stated at 5 to 20% come out yes more often than stated. The multi-label roll-up is a product of per-candidate probabilities, not a joint probability; the model never abstains; the evidence pointers are experimental. The base, Qwen3.8-27B at a pinned revision, is not included. The LoRA is applied only from the question onward, and a runtime binding refuses any engine other than the measured one; the MLX package in the repo is still v1.0. The maker disclosed three CC BY-SA documents among its 160 training documents. The third-party Decision Index (0.3 edition) lists it as "Solomon v1.1" with partial coverage.
What it decides
- choice — picks one option from a set
- noul — answers a yes/no question with one probability
- classify — assigns a category from a fixed taxonomy
At a glance
| Parameters | 27B |
| Base model | Qwen/Qwen3.8-27B |
| Maker | Doccy Health |
| Released | 2026-09-21 |
| License | apache-2.0 |
| Reported accuracy | 88.0% |
Get the weights
pip install systemonemodels
systemone pull doccy-health/solomon
The files are served from the maker's Hugging Face repository, DoccyHealth/Solomon, and verified against the checksums recorded here.